FEATURE Dynamic Testing of Civil Engineering Structures Series
PRACTICAL GUIDELINES FOR THE NATURAL EXCITATION
TECHNIQUE (NExT) AND THE EIGENSYSTEM REALIZATION
ALGORITHM (ERA) FOR MODAL IDENTIFICATION USING
AMBIENT VIBRATION
by J.M. Caicedo
odal identification methodologies are used to
M determine the dynamic characteristics of sys-
tems. In structural engineering, these dynamic
parameters play an important role in the under-
standing of the dynamic behavior of bridges, building, and
other structures. Modal identification has been used for the
verification of structural designs,1 – 4 the study of the reli-
ability of a structure to extreme dynamic loads such as
earthquakes or hurricanes,5 and it has been the focus of
a large number of structural health monitoring (SHM)6 – 11
and model updating techniques.12 – 15 These methodologies
can be classified in input–output and output-only meth-
ods. Input–output methodologies require measuring both,
the force exciting the structure and the reaction of the
structure to this excitation (i.e., acceleration, strain, dis-
placement, etc.) In contrast, output-only methodologies do
not require measuring the forces acting on the structure.
In this case, the excitation is assumed stochastic with some
specific characteristics.
Controlled excitation is difficult to achieve in most civil
structures because of their size and because the influence of
external loads such as wind and traffic is difficult to isolate.
In addition, structure owners are usually reluctant to close
structures for testing because of the lost of productivity
due to the structure’s down time. The George Washington
Bridge over the Hudson River in New York city, for example,
is crossed by an estimated 300,000 vehicles each day. The
closing of this structure would create major traffic problems
in Manhattan. Therefore, output-only modal identification
using ambient excitation such as traffic, or wind is preferred
and frequently the only viable option for modal identification
of civil infrastructure.
Several methodologies are available to identify modal char-
acteristics based on ambient vibration. Among the most
common methods are the stochastic subspace identifica-
tion (SSI) and the eigensystem realization algorithm (ERA)
when applied with the natural excitation technique (NExT).
These methodologies are based on the assumptions that the
structure behaves within a linear range, the structure is
time invariant, and the forces applied to the structure are
Editor’s Note: This article is part of the ongoing Feature Series on Dynamic Testing
of Civil Engineering Structures. This series covers a wide range of technologies
appropriate to civil engineering structures from both practical/technical and
analytical perspectives. Series editor: Paul Reynolds, The University of Sheffied.
J.M. Caicedo (caicedo@engr.sc.edu) is an assistant professor with the Department of
Civil and Environmental Engineering, University of South Carolina, Columbia, SC.
52 EXPERIMENTAL TECHNIQUES July/August 2011
uncorrelated with the response of the system. SSI can be
applied directly to ambient vibration records16 and have
fewer parameters to consider when compared with other
methods.17 The ERA was originally deployed for impulse
excitation.18,19 NExT can be used to obtain records with
the same characteristics as free response data from ambi-
ent vibration, enabling the use of the ERA with ambient
vibration.
This paper focuses on the use of the NExT and the ERA
for output-only modal identification of civil infrastructure.
This paper is divided into two parts. First, a brief descrip-
tion of the methodologies is presented. This description is
followed by a discussion of the effect of each of the parame-
ters involved in these techniques. The goal of this paper is to
provide guidance on the parameter selection for NExT–ERA.
A complete discussion of the methodologies and comparison
with other techniques can be found elsewhere.18 – 22 The data
of the second experimental phase of the American Society of
Civil Engineers (ASCE)-SHM benchmark problem is used as
an example through this section. These tests were performed
in a four-story scaled building at the University of British
Columbia in 2002,20 and the data is available for download
at http://mase.wustl.edu/wusceel/asce.shm/.
IDENTIFICATION PROCESS
Modal identification using the NExT and ERA has three
main steps. First, the ambient excitation data is processed
with NExT to obtain a signal with the same characteristics
of a free vibration data. Parameters such as the number
of points for fast Fourier transforms (FFTs) and sampling
frequency of the data are important for this step. Then, the
ERA is used to obtain a numerical model of the system in
steady-state form based on this free vibration signal. Here,
the number of expected modes of vibration, and the number
of rows and columns of the Hankel matrix are parameters
of the algorithm that can affect the identification process.
Finally, natural frequencies, mode shapes, and damping
ratios are calculated from the identified numerical model.
Other factors related to the data acquisition and cleansing
of the data, such as aliasing and quantization problems,
will have a direct effect on the identification process. These
factors are not discussed in this paper.
Natural Excitation Technique
Natural excitation technique was first used for the modal
testing of a wind turbine under wind excitation21,22 and it
doi: 10.1111/j.1747-1567.2010.00643.x
© 2010, Society for Experimental Mechanics
NExT AND ERA FOR MODEL IDENTIFICATION USING
AMBIENT VIBRATION
has been used in a large number of studies. Consider a
multi-degree-of-freedom structure described by the differen-
tial equation of motion
Mÿ(t) + Cẏ(t) + Ky(t) = F(t) (1)
where y(t) is the vector of displacements at time t, F(t) is
the vector of forces at each time interval, M, C, and K are
the mass, damping, and stiffness matrices respectively, and
( · ) indicates the derivative with respect to time. Assuming
that the excitation is a white noise input, it can be shown
that the cross-correlation function between the vector of
acceleration with a reference acceleration signal Rÿ,ÿi (τ ) is
MR̈ÿ,ÿi (τ ) + CṘÿ,ÿi (τ ) + KRÿ,ÿi (τ ) = 0 (2)
This equation has the same form as the homogeneous
differential equation of motion and implies that the cross-
correlation function Rÿ,ÿi (τ ) has the same characteristics
as a free vibration signal from the structure. The cross-
correlation function can be estimated using two different
methods: (1) a direct procedure and (2) via FFTs.23 The direct
procedure uses time domain data while the estimation via
FFTs requires calculating spectral density functions. Consid-
ering discrete records of acceleration, the cross-correlation
function between two channels of acceleration i and j is
estimated using the direct approach using the equation:
1 
L−k
R̈ÿi ,ÿj (kt) = ÿi (l)ÿi (l + k)< (3)
N −k
l=1
where t and L are the time step and the total number
of points of the acceleration record respectively. The sec-
ond method to estimate the correlation function requires the
estimation of the cross-spectral density function and taking
its inverse Fourier transform.23 The power spectral density
estimation can easily be performed by Welch’s method23 and
it is available in mathematical programs such as Matlab. In
this process, the time record is divided in data blocks, and
the power spectral density is calculated by averaging the
squared magnitudes of the spectral FFT of each data block.
The analyst should determine the number of points for the
FFT, number of data points overlapping between each block,
and the record length or number of blocks for the calcula-
tions. In addition, the analyst needs to decide what reference
channel (or channels) should be used to calculate the cross-
correlation function, and the type of tapering function (i.e.,
Hanning or Hamming) used for each data block. It is impor-
tant to mention that the number of points of the resultant
cross-correlation functions is equal to the number of points
used for the FFT, and its sampling frequency is equal to the
sampling frequency of the data record.
Eigensystem Realization Algorithm
Considering the state-space representation for a discrete
system
x(k + 1) = Ax(k) + Bu(k) (4)
y(k) = Cx(k) + Du(k) (5)
where x(k) is the vector of states, u is the vector of system
inputs (i.e., forces applied to the structure) and y is the vec-
tor of system outputs (i.e., acceleration measurements used
modal identification) at the k-th step, and A, B, C, and D are
the discrete-time state-space matrices.24
The eigensystem realization algorithm uses the principles of
minimum realization to obtain a state-space representation
of the structure. A realization is the estimation of the system
matrices A, B, and C from the response of the structure.
There are an infinite number of matrices A, B, C, and D,
each of different dimensions, that can be used to describe the
input/output relationship of the structure. However, we are
interested in the case with the smallest number of states.
This realization is called a minimum realization. The eigen-
values of A are complex conjugates and are called the poles
of the system. Each complex conjugate pair corresponds to
one mode of vibration and can be used to determine the nat-
ural frequency and the damping ratio of the structure. The
imaginary
 part represents the damped natural frequency
(±ωn 1 − ξn2 ) and the real part corresponds to −ξn · ωn ,
where ωn stands for the natural frequency and ξn for the
damping ratio of the n-th mode of vibration.
When the ERA is used in conjunction with NExT, matrices B
and D cannot be calculated because the input to the system
is not known. Nevertheless, the modal parameters of the
system can be calculated from the matrices A and C. The
first step in applying ERA is forming the Hankel matrix18,19
⎡ ⎤
y(k + 1) y(k + 2) ··· y(k + m)
⎢ y(k + 2) y(k + 3) · · · y(k + m + 1) ⎥
⎢ ⎥
H(k) = ⎢ .. ⎥
⎣ . ⎦
y(k + n) y(k + n + 1) ··· y(k + m + n)
(6)
where the number of rows and columns of the Hankel matrix
is n × N and M respectively, and N is the number of sensors
available for identification. The second step is to perform the
singular value decomposition of H(0)
H(0) = RS T (7)
where H(0) is the Hankel matrix at k = 0, R and S are m
by m and n by n orthonormal matrices respectively, and  is
a m by n matrix with nonnegative numbers in the diagonal.
Under ideal conditions, the matrix  will be
g 0
= (8)
0 0
where  g is a g by g matrix and g is the system order or num-
ber of poles. In reality, the diagonal terms of the matrix  are
nonzero produced by noise in the data acquisition process and
numerical truncation. A minimum realization is obtained by
eliminating the smaller singular values, resulting in a min-
imum order system that represents the structure. It can be
shown that the system matrices A and C can be calculated as:
A =  −1/2 RT H(1)S −1/2 (9)
C = ET R 1/2 (10)
where the small singular values have been eliminated from
the matrix  and E = [ I 0 ] with I being an m by m identity
matrix and 0 a matrix of appropriated dimensions. For the
July/August 2011 EXPERIMENTAL TECHNIQUES 53
NExT AND ERA FOR MODEL IDENTIFICATION USING
AMBIENT VIBRATION
Fig. 1: Benchmark structure
successful implementation of the ERA, the analyst needs to
decide the number of rows and columns of the Hankel matrix
H, and the expected number of natural frequencies (or poles).
In addition, the analyst should be able to judge the results
of the identification based on a particular set of metrics.
The following section provides some practical guidelines to
perform the identification and evaluate the estimated modal
parameters.
PARAMETER SETUP: PRACTICAL
GUIDELINES
The successful implementation of NExT and ERA for modal
identification requires the analyst to setup over five different
Fig. 2: Representative acceleration record
54 EXPERIMENTAL TECHNIQUES July/August 2011
parameters. The number of points of the FFT, and the refer-
ence channel used to calculate the cross-correlation function
are the two main parameters for NExT. The ERA has
three parameters to be determined: the number of rows
and columns of the Hankel matrix (m and n), and the num-
ber of poles to identify. In addition, the length and sampling
frequency of the time domain records affect the modal iden-
tification procedure. Usually, this is an iterative procedure
where the analyst performs the identification with a set of
parameters, analyzes the output of this process and changes
the parameters depending on the results obtained. Once the
parameters are set for a specific structure, subsequent anal-
ysis for the same structure can be performed with little or
no changes. This section describes best practices to perform
modal identification using NExT–ERA, including guidelines
to select these parameters appropriately. The records from
the experimental phase of the ASCE-SHM benchmark prob-
lem are used to illustrate the change in parameters.
Test Descriptions
Acceleration records from the experimental phase of the
International Association of Structural Control (IASC)-ASCE
structural health monitoring benchmark problem are used
to demonstrate the application of NExT and ERA. Figure 1
shows a picture of the two-bay by two-bay four-story steel
structure used for this study. Ambient, hammer, and shaker
excitations were used during the experiment. Sixteen data
channels were collected during each test, 15 corresponding
to acceleration records and one corresponding to the exci-
tation force when the shaker or hammer was used. Each
story had three accelerometers, one in the center beam mea-
suring acceleration in the east–west direction and two in
the east and west sides of the building measuring in the
north–south direction. The structure was modified in nine
different configurations by changing braces and loosening
bolts. This paper considers the acceleration records recorded
with the fully braced system under ambient excitation only.
Figure 2a shows a representative acceleration record corre-
sponding to the braced configuration of the structure with
ambient vibration. Figure 2b shows the first two moments
of the signal (mean and standard deviation) calculated in
NExT AND ERA FOR MODEL IDENTIFICATION USING
AMBIENT VIBRATION
windows of 1000 points with 500 points overlapping. The
mean and standard deviation are relatively constant, giving
an indication of the stationarity of the process. The experi-
mental records are readily available for download from the
ASCE structural health monitoring committee page. More
information about the experimental benchmark problem can
be found in Reference 25 or in the committee webpage.
Record Selection
A successful modal identification starts with understanding
the available data records. The sampling frequency, noise-to-
signal ratio, and stationarity of the records are key factors.
Here, we assume that the data records have a relatively
low noise-to-signal ratio and focus on the stationarity and
the sampling frequency of the records. Natural excitation
technique and ERA can successfully be used for modal iden-
tification using records that are not entirely stationary, but
having stationary records such as the one shown in Fig. 2
produces the best results. Records are weakly stationary if
the mean and the correlation functions do not change with
time, that is, when the mean at time μ(t) is constant, and
the correlation function Rxx (t, t + τ ) = Rxx (τ ) for any time
t. A record is strongly stationary when all possible moments
and joint moments of the signal do not change with time.26
If part of the record is stationary and part of a record is non-
stationary, such as those obtained in bridges having periods
of continuous traffic versus times of sporadic traffic, the non-
stationary part can be discarded. Before starting the modal
identification procedure, it is useful to create a few cross-
spectral density functions and determine if clean peaks are
shown within the spectrum. These peaks show the existence
of possible natural frequencies. If peaks are not found either
use a higher sampling rate to capture the dynamics of the
system, or check the sensor location or noise-to-signal ratio.
These spectral plots can also be used to identify the range of
frequencies to be identified.
Sampling Frequency, Number of Points for
FFT, Record Length, and Reference Channel
An important step in the modal identification using NExT
and ERA is to calculate a high-quality cross-correlation func-
tion. The three parameters that affect the performance of
NExT when calculating the cross-correlation function via
FFT are the record length, number of points for the FFT,
overlapping, and sampling frequency of the data record. Here
the number of points for the FFT corresponds to the window
length or data block. These parameters affect the length
and frequency content of the cross-correlation function. The
record length and the amount of overlapping between win-
dows affect the number of averages for the spectral density
function calculations and it should be considered in conjunc-
tion with the number of points to be used in the FFT. Figure 3
shows two cross-power spectral density plots. Both plots are
calculated based on the same acceleration record and using
1024 points for the FFT. The sampling frequency of the
acceleration record is 200 Hz, making one data block equal
to 5.12 s. The solid line in Fig. 3 was calculated with 60 s of
acceleration, while dashed uses 30 s. As Expected, the cross-
spectral density calculated with the longer record is smoother
than the cross-spectral density calculated with the shorter
record. Longer records are preferred for noisy data, while
Fig. 3: Cross-spectral density function with two
record lengths
relatively short records can be used for relatively clean data.
The record length should be selected based on the expected
frequency of the structure. Lower frequency structures such
as cable-stayed bridges require longer records to capture the
same number of cycles than higher frequency structures.
The sampling frequency determines the frequency range
where modal parameters are identified. Having a record with
a high sampling frequency can produce a lack of performance
in the identification of some low frequency modes, especially
when they are closely spaced. Often the data should be resam-
pled to focus the methodology in a specific frequency range.
The sample frequency should be at least twice the higher fre-
quency of interest to comply with the Nyquist criterion but
it is recommended to select a sampling frequency not much
higher than that. Anti-aliasing filters should be used when
resampling the data. Otherwise the modal identification pro-
cedure could incorrectly identify high frequency modes at a
lower frequency due to aliasing on the signals. However, it is
important to consider that anti-aliasing filters could appear
as poles on the signals due to filter’s dynamic characteristics.
Considering the cross-spectral density shown in Fig. 3, one
can argue that most of the data of interest is below 40 Hz,
and therefore we can resample the data to 80 Hz.
The sampling frequency and the number of points for the
FFT affect the length of the cross-correlation record as the
sampling frequency of the record is equal to the sampling fre-
quency of the cross-correlation, and the number of points in
the cross-correlation function is equal to the number of points
of the FFT. Selecting high sampling frequencies and keeping
the number of points of the FFT constant can produce either
too short or long cross-correlation records. Figure 4 shows
the free vibration record calculated with the original record
(Fig. 2) sampled at 200 Hz and with the record resampled at
80 Hz. Both cross-correlation functions were calculated with
1024 points in the FFT. The first cross-correlation is just over
2.5 s while the second one is over 6 s. The free decay is more
evident in the cross correlation calculated with 80 Hz because
it goes further in time even though it has the same number of
July/August 2011 EXPERIMENTAL TECHNIQUES 55
NExT AND ERA FOR MODEL IDENTIFICATION USING
AMBIENT VIBRATION
Fig. 4: Cross-correlation functions
points. Even though short cross-correlations not displaying a
clear visible free decay can be successfully used for identifica-
tion, having long records of the correlation allows the analyst
to judge the quality of the correlation estimate visually.
Natural excitation technique requires a reference channel
to calculate the cross-correlation function. One option is to
choose a reference channel with high amplitude, low noise-
to-signal ratio, and far from a node of vibration of the modes
of interest. Using signals collected from sensors close to a
node of vibration of a particular mode can create challenges
in the identification this mode. An alternative to select a sin-
gle reference channel is to perform the modal identification
several times with different reference channels each time.
This allows the creation of a diagram similar to stabilization
diagrams,27 allowing a more thorough modal identification.
In addition, the ERA can support data using several refer-
ence channels. In this case the Hankel matrix entries become
matrices instead of vectors. This method is explained in detail
in the following section.
Rows and Columns of the Hankel Matrix
The number of rows and columns of the Hankel matrix are
two of the main parameters of the ERA. The number of rows
and columns is chosen based on the number of expected
natural frequencies on the system. If the analyst is familiar
with the system dynamics or has a numerical model of the
structure, a first approximation to the number of natural
frequencies can be guessed based on experience or a prelimi-
nary numerical model. If the dynamics of the structure is not
known, a first approximation to the number of modes can be
obtained by counting the peaks of the cross-spectral density
function.
A rule of thumb for the number of columns of the Hankel
matrix is to use four times the number of expected modes
(twice the number of expected poles). It is important to men-
tion that forcing a low-rank Hankel matrix could lead to
missing some physical modes of vibration but this is usually
a good starting point. The number of rows is a set based on
56 EXPERIMENTAL TECHNIQUES July/August 2011
the number of points available in the cross-spectral density
function. The goal is to use as much data from the spectral
density function as possible without including noisy signals
found at the end of the cross-correlation function. Consid-
ering, for example, the cross-correlation function shown in
Fig. 4b, the first 5 s of data, where a clear decay is present,
are used to form the Hankel matrix. A few cycles of vibration
should be used to ensure that the modal identification is per-
formed successfully. Therefore, the cross-correlation record
should be longer for structures with low natural frequencies.
Model Order and Stabilization Diagrams
Specifying the correct number of poles, or model order, is
probably the most important step in the modal identification
procedure. If the model order is too high, fictitious modes
of vibration can be included in the identified results. If the
model order is too small, some of the modal parameters might
not be identified. Stabilization diagrams are an effective tool
to determine the correct number of poles for the ERA.28 – 30
The idea behind stabilization diagrams is to repeat the iden-
tification process with a different number of poles each time.
Stable poles should remain constant for all or most of the
iterations. Figure 5a shows an example of a stabilization
diagram.
Stabilization diagrams are also used to explore the use of
different reference channels for the calculation of the cross-
spectral density functions. Different modes can be identified
when different reference channels are selected for NExT,
and therefore the final modal identification could be a mix of
modes identified with different reference channels. Figure 5a
shows the stabilization diagrams for two reference channels.
The dots indicate the identified poles using the acceleration
record of the 4th floor in the east–west direction as a
reference channel for NExT and the circles indicate the
poles identified when the 4th floor, north–south acceleration
record was used. In this particular structure, a possible
mode at 23 Hz is only observed when using the north–south
acceleration record as a reference channel.
NExT AND ERA FOR MODEL IDENTIFICATION USING
AMBIENT VIBRATION

Fig. 5: Stabilization diagrams

MAC and Damping Filter
An additional step can be used to evaluate the identified
parameters from the stabilization diagram. The damping
ratio and modal assurance criterion (MAC) values can be
used to automatically identify stable modes in the stabi-
lization diagram.27 Modes with damping ratios lower or
higher than certain threshold limit can be automatically
discarded. These modes commonly correspond to numeri-
cal modes. The damping estimation in these techniques has
a larger variation when compared with the estimation of
the natural frequencies, and mode shapes and the selected
threshold limits should reflect this variation. In addition,
the correlation between the identified mode shapes at a spe-
cific frequency can be compared using the MAC. The MAC
value is metric used to determine the similarity (linearity)
of one-mode shape when compared with another reference
mode. Modal assurance criterion values range from zero to
one, zero corresponding to no consistent mode shapes, and
one corresponding to consistent shapes.31 The MAC value is
calculated using the equation
unstable modes can be discarded. As a rule of thumb, MAC
values higher than 0.95 indicate a good correlation between
two mode shapes. Mode shapes with lower MAC values can
be considered fictitious modes and discarded from the iden-
tification. Figure 5b shows the same stabilization diagram
presented in Fig. 5a but including only the modes with MAC
values higher than 0.95 and damping ratios lower than 5%.
The steps to obtain this graph are the following: (1) discard all
modes with more than 5% damping, (2) select the first identi-
fied natural frequency, (3) search for modes with frequencies
within 0.005 Hz of the selected frequency, (4) calculate the
MAC between these modes, (5) if the number of modes with
MAC values greater then 0.95 is more than 20, accept the
selected modes as a true mode of vibration, (6) identify the
following unknown mode of vibration and repeat the pro-
cess starting at (3). Six natural frequencies could be clearly
identified after this process corresponding to the first three
bending mode in the north–south direction (7.49 and 19.9 Hz
and 25.45 Hz, respectively), the first two bending modes in
the east–west direction (7.77 and 19.90 Hz, respectively) and
a rotational mode (14.47 Hz). Note that the frequencies of the

2 first two modes of the structure are closely spaced but the
n
j =1 φa,j φb,j methodology was successful in identifying them.
MACa,b = 
 
 (11)
n 2 n 2
j =1 (φa,j ) j =1 (φb,j )
where φa,j is the j -th coordinate of the mode shape a. The
MAC is used here to determine stable mode shapes on the
stabilization diagram. As the real mode shape is not known,
MAC is used to compare the identified modes in each iteration
of the stabilization diagram at a particular frequency. Any
mode shape is selected as a reference mode and compared
with all other modes at the same frequency. The result of this
operation is a vector of MAC indicating which modes are sim-
ilar to the reference shape. A MAC matrix can be calculated
selecting each mode identified as a reference mode. Here each
column corresponds to different reference mode shapes. The
diagonal of the matrix should be equal to one since it is the
MAC calculated using one-mode shape. Columns with a high
mean value indicate that the selected reference shape for that
particular column has been identified several times in the sta-
bilization diagram. Once a stable reference shape is selected
CONCLUSIONS
This paper discusses the steps and parameters to perform
an effective modal identification using the NExT and the
ERA. The steps for a successful modal identification can
be summarized in: (1) data record selection, (2) calculation of
cross-correlation functions, (3) identification of system matri-
ces, and (4) calculation of modal parameters. The record
length, number of points for the FFT and the record
sampling frequency directly affect the calculations of the
cross-correlation function. The number of rows and columns
for the Hankel matrix as well as the number of expected
modes are important parameters for the identification of the
system matrices using the ERA. The use of stabilization
diagrams, MAC values, and damping ratios are used to dif-
ferentiate numerical modes from real modes. Acceleration
records from the IASC-ASCE experimental structural health
July/August 2011 EXPERIMENTAL TECHNIQUES 57
NExT AND ERA FOR MODEL IDENTIFICATION USING
AMBIENT VIBRATION
monitoring benchmark problem were used to illustrate the
use of the methodology. Six natural frequencies, including
two closely spaced modes were successfully detected.
References
1. Hui, M.C.H., Ding, Q.S., and Xu, Y.L., ‘‘Flutter Analysis of
Stonecutters Bridge,’’ Wind and Structures 9:125–146 (2006).
2. Zivanovic, S., Pavic, A., and Reynolds, P., ‘‘Modal Testing and
FE Model Tuning of a Lively Footbridge Structure,’’ Engineering
Structures 28:857–868 (2006).
3. Ren, W.X., Peng, X.L., and Lin, Y.Q., ‘‘Experimental and
Analytical Studies on Dynamic Characteristics of a Large Span
Cable-stayed Bridge,’’ Engineering Structures 27:535–548 (2005).
4. Ni, Y.Q., Wang, J.Y., and Lo, L.C., ‘‘Influence of Stabilizing
Cables on Seismic Response of a Multispan Cable-stayed Bridge,’’,
Computer-Aided Civil and Infrastructure Engineering 20:142–153
(2005).
5. Siringoringo, D.M., and Fujino, Y., ‘‘Observed Dynamic Per-
formance of the Yokohama-Bay Bridge from System Identification
Using Seismic Records,’’ Structural Control and Health Monitoring
13:226–244 (2006).
6. Doebling, S.W., Farrar, C.R., Prime, M.B., and Shevitz, D.W.,
Damage Identification and Health Monitoring of Structural and
Mechanical Systems from Changes in Their Vibration Character-
istics: A Literature Review, Los Alamos National Laboratory, Los
Alamos, New Mexico, LA -13070-MS (1996).
7. Sohn, H., Farrar, C.R., Hemez, F., Shunk, D.D., Stine-
mates, D.W., and Nadler, B.R., A Review of Structural Health Mon-
itoring Literature: 1996–2001, National Laboratory, Los Alamos,
LA-13976-MS (2003).
8. Ching, J., and Beck, J.L., ‘‘Bayesian Analysis of the Phase II
IASC-ASCE Structural Health Monitoring Experimental Bench-
mark Data,’’ Journal of Engineering Mechanics 130:1233–1244
(2004).
9. Caicedo, J.M., and Dyke, S.J., ‘‘Experimental Validation of
Structural Health Monitoring for Flexible Bridge Structures,’’
Journal of Structural Control and Monitoring 12:425–443 (2005).
10. Caicedo, J.M., Dyke, S.J., and Johnson, E.A., ‘‘Natural
Excitation Technique and Eigensystem Realization Algorithm for
Phase I of the IASC-ASCE Benchmark Problem: Simulated Data,’’
Journal of Engineering Mechanics 130:49–60 (2004).
11. Giraldo, D., Dyke, S.J., and Caicedo, J.M., ‘‘Damage Detection
Accommodating Varying Environmental Conditions,’’ Structural
Health Monitoring 5:155–172 (2006).
12. Brownjohn, J.M.W., and Xia, P.-Q., ‘‘Dynamic Assessment
of Curved Cable-Stayed Bridge by Model Updating,’’ Journal of
Structural Engineering 126:252–260 (2000).
13. Katafygiotis, L.S., and Yuen, K.-V., ‘‘Bayesian Spectral Den-
sity Approach for Modal Updating Using Ambient Data,’’ Earth-
quake Engineering and Structural Dynamics 30:1103–1123 (2001).
14. Beck, J.L., and Katafygiotis, L.S., ‘‘Updating Models and
Their Uncertainties. I: Bayesian Statistical Framework,’’ Journal of
Engineering Mechanics 124:455–461 (1998).
58 EXPERIMENTAL TECHNIQUES July/August 2011
15. Katafygiotis, L.S., and Beck, J.L., ‘‘Updating Models and
Their Uncertainties. II: Model Identifiability,’’ Journal of Engi-
neering Mechanics 124:463–467 (1998).
16. Van Overschee, P., and De Moor, B., Subspace Identification
for Linear Systems: Theory, Implementation, Applications, Kluwer
Academic Publishers, Boston (1996).
17. Giraldo, D.F., Song, W., Dyke, S.J., and Caicedo, J.M., ‘‘Modal
Identification through Ambient Vibration: Comparative Study,’’
Journal of Engineering Mechanics 135(8): 759–770 (2009).
18. Juang, J.-N., and Pappa, R.S., ‘‘An Eigensystem Realization
Algorithm for Modal Parameter Identification and Model Reduc-
tion,’’ Journal of Guidance 8:620–627 (1985).
19. Juang, J.-N., and Pappa, R.S., ‘‘Effects of Noise on Modal
Parameters Identified by the Eigensystem Realization Algorithm,’’
Journal of Guidance 9:294–303 (1986).
20. Dyke, S.J., Bernal, D., Beck, J.L., and Ventura, C.E., ‘‘Exper-
imental Phase of the Structural Health Monitoring Benchmark
Problem,’’ 16th Engineering Mechanics Conference, Reston, Virginia
(2003).
21. James, G.H., Carne, T.G., and Lauffer, J.P., The Natural
Excitation Technique (NExT) for Modal Parameter Extraction From
Operating Wind Turbines, Sandia National Laboratories, Albu-
querque, NM and Livermore, CA, SAND92-1666 (1993).
22. James, G.H., Carne, T.G., and Lauffer, J.P., ‘‘The Natural
Excitation Technique (NExT) for Modal Parameter Extraction from
Operating Structures,’’ Modal Analysis 10:260–277 (1995).
23. Welch, P.D., ‘‘Use of Fast Fourier Transform for Estima-
tion of Power Spectra - a Method Based on Time Averaging over
Short Modified Periodograms,’’ IEEE Transactions on Audio and
Electroacoustics 15:70–73 (1967).
24. Juang, J.-N., Applied System Identification, Prentice Hall,
Englewood Cliffs, NJ (1994).
25. Dyke, S.J., Bernal, D., Beck, J.L., and Ventura, C.E., ‘‘Exper-
imental Phase of the Structural Health Monitoring Benchmark
Problem,’’ 16th Engineering Mechanics Conference, Reston, Virginia
(2003).
26. Bendat, J.S., and Piersol, A.G., Random Data: Analysis and
Measurement Procedures, 3rd Edition, Wiley, New York (2000).
27. Giraldo, D., ‘‘A Structural Health Monitoring Framework for
Civil Structures,’’ Department of Civil Engineering, Washington
University in St. Louis, Missouri, p. 157 (2006).
28. Basseville, M., Benveniste, A., Goursat, M., Hermans, L.,
Mevel, L., and Van der Auweraer, H., ‘‘Output-only Subspace-
based Structural Identification: From Theory to Industrial Testing
Practice,’’ Journal of Dynamic Systems Measurement and Control-
Transactions of the Asme 123:668–676 (2001).
29. Allemang, R.J., and Brown, D.L., ‘‘A Unified Matrix Poly-
nomial Approach to Modal Identification,’’ Journal of Sound and
Vibration 211:301–322 (1998).
30. Yu, D.H., and Ren, W.X., ‘‘EMD-based Stochastic Subspace
Identification of Structures from Operational Vibration Measure-
ments,’’ Engineering Structures 27:1741–1751 (2005).
31. Allemang, R.J., ‘‘The Modal Assurance Criterion - Twenty
Years of Use and Abuse,’’ Journal of Sound and Vibration 37:14–23
(2003).